Microsoft AI-103 Practice Test 47
Question 460
A document is deleted for a legal retention request but its embedding remains in the index. What action is needed?
- Remove the source document record but keep its embeddings and cached passages active
- Return source-grounded citations that identify exact evidence spans
- Remove or invalidate corresponding vectors and cached excerpts during deletion handling
- Apply source ACL and tenant authorization at retrieval time
Explanation
Correct Answer
C:Derived retrieval artifacts must follow source retention and deletion requirements. The decisive difference is whether the design can remove expired source-derived vectors and snippets; the competing function serves to identify referenced document revisions.
Incorrect Answers
A:A deletion workflow must propagate to derived vectors and caches, otherwise removed content may continue to appear in retrieval.
B:Verifiable passage references let readers assess whether an answer follows the authorized documents. Here the needed outcome is to remove expired source-derived vectors and snippets; an option designed to attach traceable evidence to generated answers would not directly provide it.
D:A shared vector index must not bypass permissions present on the authoritative documents. That is relevant to prevent restricted documents crossing retrieval boundaries, but the present scenario requires an ability to remove expired source-derived vectors and snippets.
Question 461
An agent should provide an explanation of its action without revealing private internal prompts or secrets. What output should be logged?
- Publish the complete hidden system prompt and all raw tool secrets with the audit record.
- Return only a success flag without recording which approved action occurred.
- Return concise user-facing decisions with permitted evidence and structured action provenance.
- Reconstruct a fictional reasoning transcript in place of recorded events.
Explanation
Correct Answer
C:Auditable decisions do not require exposing hidden prompts, security tokens or untrusted raw tool transcripts.
Incorrect Answers
A:Internal prompts and credentials are not required for accountable user-facing explanations and may be sensitive.
B:A bare flag lacks the actor, scope, evidence and outcome necessary to explain the action.
D:Fabricated reasoning is not reliable provenance and can misrepresent the actual workflow.
Question 462
A team optimizes solely on an LLM judge's score, then discovers the judge prefers longer outputs. What evaluation improvement helps?
- Treat a higher score from the same uncalibrated language-model judge as unquestionable quality improvement.
- Calibrate evaluator scoring against labeled human-reviewed examples
- Provide traceable source references and decision-factor provenance for review
- Inspect agent trajectory and policy compliance in addition to final-answer quality
Explanation
Correct Answer
B:Human anchor examples reveal judge bias and improve confidence in automatic quality metrics. This resolves the specific need to detect evaluator preference bias, while a measure that helps evaluate whether answers actually fulfill the task addresses a different requirement.
Incorrect Answers
A:A judge's style bias can reward verbosity without improved correctness. Calibrate scores against reviewed examples and independent task criteria.
C:Auditable evidence and factor summaries explain why the system recommended an outcome without exposing private hidden reasoning. This helps when a team needs to make AI-supported decisions explainable with evidence; it does not establish how to detect evaluator preference bias.
D:A correct final sentence does not establish that intermediate tools were used safely or with authorization. The function serves to evaluate agent actions as well as final responses, whereas the actual requirement is to detect evaluator preference bias.
Question 463
A retail content pipeline accepts customer images, generates promotional media and must guard against embedded malicious text. Which THREE controls belong in it? Choose THREE.
- Apply clear provenance labeling or watermarking where appropriate
- Treat OCR-discovered text as untrusted input and constrain agent actions by policy
- Publish customer-provided logos as permitted marketing material without verifying rights or usage policy.
- Skip visual safety calibration and approve uploaded media solely when an unverified classifier returns any score.
- Treat routine promotional images as preapproved and bypass both incoming media screening and the embedded-text trust boundary.
- Screen incoming media with appropriate visual safety classification
Explanation
Correct Answers
A:Synthetic-media disclosure reduces confusion over whether visuals show real products or events.
B:Text inside pictures can carry indirect prompt injection; perception does not confer instruction authority.
F:Input moderation reduces exposure to disallowed visual content before later processing.
Incorrect Answers
C:Customer uploads do not convey permission to reuse protected marks in marketing outputs.
D:An uncalibrated, unchecked score cannot guarantee safe media intake. The pipeline needs actual screening and controls on untrusted embedded text.
E:A broad approval label does not inspect malicious customer-provided imagery or enforce policy on OCR-discovered instructions.
Question 464
A safety-critical voice workflow receives an uncertain transcription before making an external change. What safeguard fits?
- Execute a side-effecting command on a low-confidence transcript without user confirmation
- Ask for explicit confirmation of the recognized command before side effects
- Use endpoint detection and turn-taking signals before final response synthesis
- Normalize modality-specific evidence into tagged multimodal conversation turns
Explanation
Correct Answer
B:Voice recognition errors can be costly, so important tool actions require verified user intent. The practical requirement is to confirm uncertain spoken commands before actions; a solution built to handle mid-session language changes safely would not meet that same need.
Incorrect Answers
A:Speech-recognition uncertainty creates a risk of executing the wrong action. Require explicit confirmation of the interpreted command before side effects.
C:Voice interaction needs explicit speech-turn boundaries to avoid interrupting or mishearing requests. The function serves to detect completion of a spoken user turn, whereas the actual requirement is to confirm uncertain spoken commands before actions.
D:Tagged input blocks preserve source and modality while maintaining ordered task context. That capability is appropriate for integrate different input modalities in one conversation, not this separate task of finding how to confirm uncertain spoken commands before actions.
Question 465
An AI assistant paraphrases a policy without revealing where the supporting claim came from. What feature increases auditability?
- Return source-grounded citations that identify exact evidence spans
- Return uncited policy statements based only on the model's stored knowledge
- Apply source ACL and tenant authorization at retrieval time
- Remove or invalidate corresponding vectors and cached excerpts during deletion handling
Explanation
Correct Answer
A:Verifiable passage references let readers assess whether an answer follows the authorized documents. In this scenario the team must attach traceable evidence to generated answers. A feature used to identify referenced document revisions does not substitute for it.
Incorrect Answers
B:Fluent statements cannot establish which current policy passage supports the answer. Evidence-linked citations are needed for an auditable response.
C:A shared vector index must not bypass permissions present on the authoritative documents. That is relevant to prevent restricted documents crossing retrieval boundaries, but the present scenario requires an ability to attach traceable evidence to generated answers.
D:Derived retrieval artifacts must follow source retention and deletion requirements. This helps when a team needs to remove expired source-derived vectors and snippets; it does not establish how to attach traceable evidence to generated answers.
Question 466
An agent drafts code that must compile before deployment. Which reflection step is defensible?
- Accept the draft because the model privately predicts its code will compile, without running a build or test.
- Use independent source-grounded checks rather than relying solely on self-confidence
- Re-run a held-out evaluation set with measured before-and-after outcomes
- Run deterministic compilation and tests and feed failures into bounded repair
Explanation
Correct Answer
D:Executable tests provide stronger correction signals than asking the model whether code works. In this scenario the team must repair generated code from actual test failures. A feature used to audit model failures using observable evidence does not substitute for it.
Incorrect Answers
A:An internal prediction cannot replace deterministic verification. Compile the generated artifact and run focused tests, using observed failures for bounded correction.
B:Self-critique without external evidence can reinforce the same mistaken claim. Its primary purpose is to verify model assertions against independent sources; the stated constraint instead demands a way to repair generated code from actual test failures.
C:Independent examples reveal whether revision improves correctness beyond cosmetic changes. That is relevant to evaluate objective improvement after a critique pass, but the present scenario requires an ability to repair generated code from actual test failures.
Question 467
A regulated customer needs to understand which records influenced each denial recommendation. What explanation artifact should be retained?
- Use groundedness or faithfulness evaluation against the source document
- Return an unexplained numerical confidence score without retaining the source records behind the recommendation.
- Version evaluation datasets, scoring rubrics and results alongside release artifacts
- Provide traceable source references and decision-factor provenance for review
Explanation
Correct Answer
D:Auditable evidence and factor summaries explain why the system recommended an outcome without exposing private hidden reasoning. The practical requirement is to make AI-supported decisions explainable with evidence; a solution built to score source-supported generated statements would not meet that same need.
Incorrect Answers
A:A fluent summary can still invent claims; groundedness tests whether assertions are supported by the reference. That is relevant to score source-supported generated statements, but the present scenario requires an ability to make AI-supported decisions explainable with evidence.
B:A score alone cannot reconstruct a regulated denial recommendation. Preserve relevant input records, authorized evidence and decision links for review.
C:Fixed benchmarks and recorded methods make regressions reproducible across candidates. The function serves to make evaluation runs reproducible over time, whereas the actual requirement is to make AI-supported decisions explainable with evidence.
Question 468
A second model reviews a first model's summary, but both models repeat the same unsupported citation. What evaluator can catch it?
- Ask a second model to rate citation confidence without showing it the actual retrieved source passages.
- Check cited statements against the retrieved source passages
- Cap critique-and-revision passes using quality improvement thresholds
- Re-run a held-out evaluation set with measured before-and-after outcomes
Explanation
Correct Answer
B:A second opinion is not enough when both share the same fabricated evidence. This resolves the specific need to validate citations at the source level, while a measure that helps audit model failures using observable evidence addresses a different requirement.
Incorrect Answers
A:Agreement between models cannot establish that a citation supports a claim. Validate quotations and assertions against the underlying source content.
C:Bounded reflection prevents token waste when repeated attempts no longer improve output. This helps when a team needs to stop unproductive self-critique loops; it does not establish how to validate citations at the source level.
D:Independent examples reveal whether revision improves correctness beyond cosmetic changes. The function serves to evaluate objective improvement after a critique pass, whereas the actual requirement is to validate citations at the source level.
Question 469
An organization wants to compare safety across demographic subgroups while honoring privacy constraints. What evaluation design is suitable?
- Version evaluation datasets, scoring rubrics and results alongside release artifacts
- Measure answer completeness and task relevance against expected criteria
- Run privacy-preserving subgroup performance and fairness checks on appropriate test data
- Report only a single aggregate model score while withholding all privacy-safe subgroup performance results.
Explanation
Correct Answer
C:Aggregate subgroup metrics can expose disparate performance without publishing individual records. Here the relevant capability must let the team check fairness across relevant cohorts. An alternative targeting make evaluation runs reproducible over time solves another problem.
Incorrect Answers
A:Fixed benchmarks and recorded methods make regressions reproducible across candidates. This helps when a team needs to make evaluation runs reproducible over time; it does not establish how to check fairness across relevant cohorts.
B:Citation correctness is not sufficient if the response does not fulfill the information need. The function serves to evaluate whether answers actually fulfill the task, whereas the actual requirement is to check fairness across relevant cohorts.
D:An aggregate score can hide unequal outcomes across groups. Use approved privacy-preserving subgroup metrics and document sample limitations.
Question 470
An agent revises an answer endlessly with small wording changes. What control limits reflection overhead?
- Allow unlimited critique-and-rewrite cycles until the model says its answer is flawless.
- Check cited statements against the retrieved source passages
- Cap critique-and-revision passes using quality improvement thresholds
- Record observable rule violations, test inputs and trace spans rather than internal hidden reasoning
Explanation
Correct Answer
C:Bounded reflection prevents token waste when repeated attempts no longer improve output. The practical requirement is to stop unproductive self-critique loops; a solution built to repair generated code from actual test failures would not meet that same need.
Incorrect Answers
A:Unbounded reflection can waste cost and repeatedly change phrasing without improvement. Apply a stopping budget, measurable tests and a termination condition.
B:A second opinion is not enough when both share the same fabricated evidence. This helps when a team needs to validate citations at the source level; it does not establish how to stop unproductive self-critique loops.
D:External events and policy criteria can explain failures without requiring private reasoning traces. That capability is appropriate for audit model failures using observable evidence, not this separate task of finding how to stop unproductive self-critique loops.